Research On Lightweight SSD Vehicle Detection Algorithm For Mobile Computing Environment

As the number of cars in my country increases year by year,in order to ease traffic pressure and improve traffic efficiency,the construction of intelligent transportation system becomes more and more important.In recent years,due to the rapid development of Deep Neural Networks,the Object Detection algorithm has been greatly improved in terms of detection accuracy and recognition accuracy,which has become an effective tool to solve the vehicle detection problem in intelligent transportation systems.This paper takes the real-time vehicle detection in the intelligent transportation system as the research task,streamlines and improves the Single Shot Multi Box Detector(SSD)target detection algorithm in terms of the Feature Extraction Network,Activation Function,Non-Maximum Suppression algorithm,Loss Function,etc.The improved SSD vehicle detection algorithm is also weight quantized from the perspective of application deployment and migrated to terminal devices for testing.This paper aims to build a lightweight real-time detection algorithm for vehicle targets.The research work mainly includes the following contents:(1)A lightweight Feature Extraction Network based on the combination of Residual Connection and Attention Mechanism is proposed for SSD target detection algorithm.The method firstly improves the Activation Function in the Residual block and SE block into less computationally intensive h-swish and h-sigmoid Activation Functions to reduce the computational load,and the lightweight SE-Res Net20 Feature Extraction Network is proposed.Then the input image size and feature map perception field are combined to reduce the number of feature map fusion layers of the SSD algorithm,and the generation ratio of the Default box is simplified according to the vehicle shape ratio at specific angles in practical applications to further reduce the amount of operations of the vehicle detection model.(2)The original SSD target detection algorithm is improved in terms of Loss Function and Non-Maximum Suppression algorithm,which first adopts the Generalized Focal Loss Function with better performance to reduce the gap between training and inference without increasing additional computation.The NMS algorithm is replaced by the Soft-NMS algorithm to improve the detection rate of overlapping targets,and then the lightweight Mobile Netv3-small is used as the backbone Feature Extraction Network,and a layer of convolution is added to the network to increase the range of feature map perception to increase the accuracy of small target vehicle recognition.(3)The improved lightweight vehicle detection algorithm is deployed to the terminal device.Firstly,the model file and weight file trained on the host computer are converted to the IR file that can be recognized and accelerated by the Neural Computation Stick 2,then the numerical accuracy is reduced by quantization to further reduce the computation required by the vehicle detection algorithm,and finally it is migrated to the Laboratory equipment built by Raspberry Pi and Neural Compute Stick for vehicle detection task experiments.